Abstract: Distributed acoustic sensing (DAS) seismic data are commonly contaminated by complex noise whose frequency bands overlap with those of useful signals, making it difficult for conventional filtering methods to balance noise suppression and waveform preservation. To address the lack of paired noisy-clean labels in practical applications, this study proposes a Wasserstein-constrained CycleGAN (WCGAN) for DAS data denoising. The method introduces the Wasserstein distance and gradient penalty into the CycleGAN framework to improve adversarial training stability and alleviate mode collapse. During training, measured noisy DAS patches and clean patches generated by two-dimensional elastic-wave forward modeling are used to construct unpaired noisy and clean domains, respectively. Tests on the public DAS dataset from the Utah FORGE project show that WCGAN can effectively suppress high-frequency random noise, horizontal banding noise, and high-amplitude burst noise while preserving the morphology and continuity of low-frequency effective signals. The proposed method provides a stable and practical intelligent solution for DAS seismic data denoising when paired annotations are unavailable.